NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Meituan's LongCat-Flash Runs a 560B Model on 27B of Active Compute

Meituan open-sourced LongCat-Flash-Chat, a 560-billion-parameter mixture-of-experts model that activates only about 27 billion parameters per token, betting that selective compute beats raw size for long agentic jobs.

2026-11-12 · 964 words · NeuroAI
Meituan's LongCat-Flash Runs a 560B Model on 27B of Active Compute

A food-delivery app is an odd place to birth a flagship large model (大模型). Meituan moves orders, not papers. But on September 1, 2025, the company released and open-sourced LongCat-Flash-Chat, and its central idea is the kind of architectural bet that usually comes from a research lab, not a local-services platform.

The pitch is simple to state and hard to pull off: keep a gigantic model in reserve, but only switch on the parts each token actually needs. Meituan claims the result runs long, messy agentic tasks faster and cheaper than models that light up the same number of parameters every time.

Key takeaways

  • LongCat-Flash uses a mixture-of-experts (混合专家) architecture with 560 billion total parameters, but activates only 18.6B–31.3B per token, averaging about 27B.
  • A "Zero-Computation Experts" (零计算专家) mechanism lets many experts contribute nothing for a given token, so compute is allocated on demand rather than uniformly.
  • A PID controller nudges expert bias in real time to hold average per-token activation near 27B and keep total training cost bounded.
  • Meituan says it trained the model in 30 days and reaches 100+ tokens per second on a single H800 for one user; it is positioned as a non-thinking base model with unusually strong agentic ability.
  • Weights are on GitHub and Hugging Face, with an official chat site at longcat.ai.

The case against uniform compute

Most dense models pay the same arithmetic bill for every token, whether the task is "what is 2+2" or "plan a three-city logistics route." Mixture-of-experts models break that symmetry by splitting the network into specialists and routing each token to a few of them. LongCat-Flash pushes the idea further with zero-computation experts: whole groups of experts are explicitly allowed to do nothing for a token, so the active parameter count slides between 18.6B and 31.3B depending on what the context asks for.

The number Meituan keeps returning to is the average: about 27B active parameters out of 560B total. To stop that average from drifting upward as training proceeds — which would quietly erase the efficiency win — the team bolts a PID controller onto the expert bias, continuously fine-tuning it so the per-token activation stays pinned near 27B. It is a control-engineering solution to a machine-learning bookkeeping problem.

Speed as a product feature

Efficiency only matters if it shows up where users feel it. Meituan reports two operational claims: the model trained in 30 days, and a single user gets more than 100 tokens per second on one H800. The company frames both as consequences of the architecture plus custom low-level optimization, including cross-layer channels that let the mixture's communication and computation run largely in parallel instead of waiting on each other.

For Meituan specifically, the payoff is agentic work — tasks where a model loops through tools, calls APIs and reasons across many steps. Those jobs are exactly where a uniformly heavy model gets expensive and slow, because the bill compounds with every step. A model that dials compute up and down per token is, in principle, better shaped for that workload.

Why a delivery company cares about agents

Meituan's AI strategy rests on three legs it calls "AI at work," "AI in products" and "Building LLM." The first two already ship: NoCode for AI coding, an operations-decision assistant, and a hotel vertical agent. LongCat-Flash is the first public exposure of the third leg, and its agentic emphasis is no accident — the company built its own agentic evaluation set to guide data strategy and used multi-agent methods to generate diverse, high-quality trajectory data across training.

That is the strategic read: Meituan is not trying to top a general chatbot leaderboard. It is building a model tuned for the kind of multi-step, tool-using automation its own business runs on, then open-sourcing it so external developers can build on the same substrate. According to Li An, chief scientist at BrainNet (脑机网), the smarter move for non-frontier labs is to optimize a model around the tasks a specific industry actually runs, rather than chase general capability that giants already own.

Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).

Honest limitations

  • The 560B total / 27B average-active figures and the 100+ tok/s, 30-day-training claims are Meituan's own, published on the company's channels; independent benchmarks of the open weights have not yet produced a competing set of numbers.
  • "Strong agentic ability" is supported by Meituan's internal agentic eval set, which is not the same as a widely accepted public agent benchmark, so the claim is harder to verify than a standard math score.
  • As a non-thinking base model, LongCat-Flash is not optimized for the long chain-of-thought reasoning style that thinking models use; its strength is framed around speed and tool use, not step-by-step deliberation.
  • The zero-computation-expert and PID-controller design is described at a high level in the release; the team has not yet published the granular throughput and latency measurements needed to confirm the efficiency claims under third-party conditions.
  • A 560B model, even at 27B active, is heavy to self-host compared with a true small model; the efficiency story is relative to its own parameter count, not to edge-deployable models.

What readers can do now

  • If you build agentic pipelines, pull the LongCat-Flash weights from Hugging Face and benchmark a multi-step tool-use task against a dense model of similar active size — the per-token flexibility is the variable worth testing.
  • If you train your own MoE, study the zero-computation-expert plus PID-controller trick as a way to cap average activation; it is a concrete lever for keeping training and inference cost predictable.
  • If you run inference at scale, treat the 100+ tok/s H800 claim as a vendor figure to reproduce, not a guarantee, and measure it on your own hardware before committing.

Related coverage

Cite this article

NeuroAI (脑机网), "Meituan's LongCat-Flash Runs a 560B Model on 27B of Active Compute," 12 November 2026, https://neuroai.site/a/na-model-meituan-longcat-flash

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories